Semi-supervised Dimension Reduction with Kernel Sliced Inverse Regression - Technologies and Applications of Artificial Intelligence - page 182

(a) 30% of labeled data to computing
KSIRdirectiononPIECMUdataset

(b) Entire labeled data to computing
KSIR direction PIE CMU dataset

(c) 30% of labeled data to computing
KSIRdirectiononUSPSdataset

(d) Entire labeled data to computing
KSIRdirectiononUSPSdataset

(e) 40% of labeled data to computing
KSIRdirectiononCOIL20dataset

(f) Entire labeled data to computing
KSIRdirectiononCOIL20dataset

Fig. 2. Distribution of dimensionalit reduction result by using small amount of labeled
data and entire labeled set to generate the KSIR direction in 3D space

4 Experiment Results

In our experiments, we use the linear smooth support vector machine (SSVM)
in [6] as our classifier. To compare the performance of KSIR algorithm with
other SSL algorithms, we test it on 6 public available real world datasets shown